| name | breakout |
| description | Raise the ceiling, don't climb toward it. When a target is nearly hit or a metric has plateaued, question the target itself: name the binding constraint, separate the physics floor from the assumed floor, design the regime change that makes 10x reachable, commit through the valley. Triggers: 'breakout', '10x this', 'raise the ceiling', 'why are we capped', 'think way bigger', 'this metric is a cage'. |
Breakout
Every other skill optimizes toward a target. This one operates on the target. Use it when the honest next move isn't a better point under the ceiling — it's a higher ceiling.
/evolve tunes parameters. /pursue redesigns the architecture to hit a given target. /breakout questions the target and changes the regime so a different, higher target becomes reachable. It does not build the new system — it sets the raised target and dispatches /pursue or /meta-harness to build it.
Shared conventions in _common.md.
When to use
| Signal | Skill |
|---|
| "Make this metric go up" | /evolve |
| "This architecture is wrong" | /pursue |
| "We're near the target and it feels small" | /breakout |
| "3 pursue/meta-harness cycles, <2% each" | /breakout — the ceiling is the target, not the code |
| "Why can't this be 10x?" | /breakout |
Do not use it to dodge honest hill-climbing. If the metric is still moving on real changes, keep /evolve-ing. Breakout is for when the ceiling is real, not when the climb is merely hard.
Procedure
- Name the binding constraint — as a mechanism, not a number. Not "we're at 0.87." "We cap at ~0.87 because retrieval re-embeds per request instead of amortizing, so recall is bounded by latency budget." If you can't name the mechanism, you're not ready to break out — go measure (
/evolve, the ground-truth harness).
- Separate the physics floor from the assumed floor. Physics floor = irreducible (security, information theory, speed of light, a real SLA). Assumed floor = how it's always been done, the metric we happened to pick, a constraint no one has retested. Attack only the assumed floor. Naming which is which is the core analytical work.
- Design the regime change. The lever is usually not the artifact. Change the metric (the current one may be the cage), the problem (eliminate the work instead of speeding it), the constraint (retest the "can't"), or the substrate (different foundation). One regime change, stated as a thesis with a falsifiable payoff.
- Set the raised target + the smallest proof the regime is reachable. A number that would be absurd under the current regime and natural under the new one, plus the cheapest experiment that shows the new regime is real (not that it's finished).
- Commit through the valley. A regime change usually regresses the old metric before it clears it. Authorize that explicitly, pair with
/dont-collapse-the-architecture so early marginal evidence doesn't kill the bet, and set the bootstrap-CI gate (_common.md) for when the bet is finally judged.
- Externalize the bet, then dispatch. Write the thesis, raised target, valley budget, and kill-condition to
.evolve/pursuits/<date>-breakout-<slug>.md. Hand the build to /pursue (single track) or /multi-pursue (independent tracks); /evolve tunes once inside the new regime.
Rules
- Optimize the ceiling, not the point under it. A 3% win that caps you forever loses to a regime change that unlocks 10x.
- If the metric can't 10x, the metric is the cage — change the metric. Then re-baseline honestly against the new one.
- Attack the assumed floor; respect the physics floor. Confusing the two is how breakout becomes either cowardice or fantasy.
- A raised target with no named mechanism is a wish. Every breakout thesis names the constraint it dissolves.
- Endurance is a state property, not a willpower property. Externalize the bet so completely that a fresh agent resumes with zero loss — then round-count stops mattering and the only stop conditions are the bet resolved or a real wall. (This is why the deterministic loop matters: you can only refuse to stop when stopping is free.)
- The valley is the price of the peak. Don't judge a regime change on week-one numbers; judge it at the gate you set in step 5.
Use references/full-reference.md for regime-change patterns, the ceiling-analysis worksheet, and the floor-separation examples.
Then consider
pursue / multi-pursue — build the regime you just defined.dont-collapse-the-architecture — the moment the first measurement comes back marginal.evolve — tune once the new regime is standing and the metric moves again.autopsy — if the breakout thesis dies, root-cause whether the regime was wrong or the proof was.
